When different training objectives pull model updates in opposing directions, causing optimization to fail or degrade.